Comparative Analysis of Deep Learning Models for Facial Emotion Recognitio

Abstract

Facial Expression Recognition (FER) plays a vital role in enabling emotionaware computing, with applications spanning healthcare, driver monitoring, human–computer interaction, online education, and entertainment. Despite significant progress, existing FER systems often struggle in real-world scenarios due to variations in lighting, occlusion, head pose, subtle expressions, and class imbalance. To address these challenges, this thesis investigates both traditional and deep learning-based approaches, followed by the development of a novel hybrid ensemble model. The research utilized two widely recognized benchmark datasets, CK+ and FER2013 Cleaned, to train and evaluate multiple pretrained transfer learning models, including VGG16, VGG19, MobileNetV2, InceptionV3, and DenseNet169. Experimental analysis revealed the strengths and weaknesses of these architectures in handling complex FER tasks. Building on these insights, a feature-level ensemble framework was proposed, integrating DenseNet169 and InceptionV3. The features extracted from both backbones were concatenated and passed through a custom multi-layer perceptron (MLP) classifier to leverage complementary global and local feature representations. The proposed Ensemble Model achieved 99.85% accuracy on CK+ and 99.34% on FER2013 Cleaned, outperforming all individual transfer learning models. These results demonstrate the effectiveness of ensemble deep learning in enhancing recognition accuracy, robustness, and scalability. While certain limitations remain, such as dataset diversity and computational cost, the findings highlight the potential of the proposed model for real-world FER applications. Future work will focus on lightweight optimization, explainable AI integration, and deployment in real-time interactive systems.